State of Robot Learning, Q1 2026: What Shipped, What Stalled

1. Frameworks: the boring infrastructure finally got boring

A year ago, picking a training framework meant picking a research group. In 2026 it has started to mean picking a contract: you commit to an API, you commit to a dataset format, and you commit to being a decent upstream citizen. Three frameworks now carry meaningful load across our own lab and the teams we advise.

LeRobot

Hugging Face's LeRobot is the closest the field has to a default. The Q1 releases focused on stability rather than surface area: better checkpoint/config provenance, cleaner dataset v2 conversions, and a policy registry that finally lets you swap ACT, Diffusion Policy, VQ-BeT, and pi0-flavored architectures without rewriting training loops. The library's opinionated dataset format has become the de facto interchange standard across small labs, and our LeRobot framework getting-started guide remains the fastest path in. The remaining rough edges are on the inference side: online policy serving on a robot still involves a lot of glue code.

Robosuite, Isaac Lab, and the sim stack

On the simulation side, Isaac Lab consolidated much of what used to be fragmented across Orbit and Isaac Gym, and Robosuite continues to be the cleanest entry point for learning research that does not need GPU-accelerated physics. We cover the tradeoffs at length in NVIDIA Isaac Lab getting started. For teams choosing between them, the deciding question remains the same: are you training thousands of environments in parallel for reinforcement learning, or are you iterating on a single manipulation task where the bottleneck is scene authoring time?

ROS 2 and the middleware layer

ROS 2 Jazzy is now the reference runtime for the majority of research arms we see shipped. The headline for Q1 was not a feature release but the degree to which learning-time tooling (LeRobot, Robosuite wrappers, teleoperation agents) now ships with ROS 2 bridges out of the box. If you are standing up a lab from scratch, our setup guides and hardware tutorials are a reasonable starting point.

2. Foundation models: the quiet consolidation of VLAs

The vision-language-action (VLA) category is where the most interesting practitioner-facing progress happened in Q1. The catalogued public models we track in /vla-models/ grew, but more importantly the usability bar moved.

OpenVLA

OpenVLA remains the most widely fine-tuned open VLA in our extended network. The combination of a 7B-parameter base with LoRA adapters brought fine-tune runs into single-GPU territory for most manipulation tasks with a few hundred episodes of demonstration data. The Q1 iteration improvements we see in the wild are mostly on the data side: better dataset curation, better action-space normalization, better handling of bimanual action heads. We discuss the budget math for fine-tuning in Scaling VLA Training on a Budget.

Octo

Octo occupies a different point in the Pareto frontier: smaller, faster to fine-tune, and more forgiving of heterogeneous datasets. For teams with small episode budgets and a single arm, Octo often wins on dollars-per-successful-evaluation even when OpenVLA wins on raw peak success rate. The practical heuristic we use: if you have fewer than ~500 episodes and one embodiment, start with Octo; above that, revisit OpenVLA.

RT-X and the cross-embodiment legacy

Google's RT-X line produced less new public substance in Q1 than some of its competitors, but its influence is everywhere. The Open X-Embodiment dataset remains the backbone against which most new VLA architectures pre-train or co-train. Our Open X-Embodiment explainer covers the dataset and how to contribute. The important Q1 development was not a new RT-X paper but the number of new labs publishing checkpoints trained on subsets of OXE with transparent recipes — the cross-embodiment idea has escaped its parent lab.

Diffusion Policy

Diffusion Policy has moved from "interesting research direction" to "default choice for contact-rich manipulation with a few hundred demos." The Q1 refresh of the reference implementation cleaned up training ergonomics and added better defaults for observation horizons. For teams choosing between ACT and Diffusion Policy, we maintain a practitioner comparison at ACT vs Diffusion Policy: when to use which.

3. Humanoid hardware: launches landed, reality checked

Q1 2026 was the quarter in which the humanoid hardware story stopped being a promise and started being a loading-dock logistics problem. Four platforms we track closely crossed meaningful delivery thresholds.

Unitree G1

Unitree's G1 kept its reputation as the workhorse research humanoid: sub-$20K entry configurations, usable SDK, a growing population of public checkpoints tuned for it, and a lead time measured in weeks rather than quarters. The tradeoffs are well-documented in our Unitree G1 review. Q1's incremental improvement was on the arms: payload, repeatability, and torque headroom all moved in the right direction, though the gripper remains the weakest link for serious manipulation work.

Booster T1

Booster Robotics graduated the T1 from early-access to general research availability in Q1. The T1's story is compliance and safety margins — a design philosophy aimed at human-proximate deployments rather than caged industrial work. Early adopters report that the control stack is less mature than Unitree's but that locomotion stability in cluttered indoor environments is strong.

EngineAI SE01

EngineAI's SE01 is the most interesting wildcard on our radar: a lower-cost locomotion-focused humanoid that arrived with competent out-of-the-box walking and a deliberately minimalist manipulation story. For labs whose near-term research agenda is whole-body control, locomotion policies, or bipedal benchmarks, SE01 materially changed the price floor in Q1.

Fourier GR-2

Fourier Intelligence's GR-2 sits at the higher end of the research price band. The Q1 shipments we saw were mostly to universities and corporate R&D labs in APAC and Europe. The story we hear consistently is that GR-2's hardware is impressive, but that lead times and spare-parts availability remain issues — a topic we take apart in Humanoid Supply Chain Reality Check 2026.

For a side-by-side on specs and price, our humanoid robot comparison 2026 is kept current with Q1 data points, and the compare tool lets you narrow by payload, reach, and price.

4. Funding: the capital is still flowing, the revenue is not

The capital story in robot learning remained extraordinary by any normal technology-industry standard in Q1 2026. The humanoid mega-rounds we tracked in 2024 and 2025 continued; several VLA and data-layer companies closed Series A and B rounds. What did not change is revenue concentration: outside industrial mobile robots and a handful of enterprise humanoid pilots, commercial revenue from general-purpose learning-based robots is still a rounding error against the capital raised.

This matters for practitioners because it sets the horizon for when closed platforms will need to either show real deployment volume or open up. Our read: 2026 is the year in which investors will start asking harder questions about unit economics. Teams building their own stacks on open datasets and open VLA checkpoints reduce their exposure to that cycle.

5. Open-source vs closed: the gap is widening, not closing

A year ago it looked like the open ecosystem (LeRobot, OpenVLA, Octo, OXE) would catch up to the closed frontier models within a year. In Q1 2026 we are calling it the other way. The open stack is more usable than ever, but the closed frontier is also moving — and the closed-side gap is increasingly about data scale rather than architecture. The short version: a lab with an open stack and 2,000 well-curated demos can match closed-model performance on narrow tasks; matching closed-model generalization across embodiments and tasks still requires either closed data licenses or an industrial-scale teleoperation operation. We cover the data-ops side in The Teleoperation Data Quality Checklist.

What did stall: the hope that a single "open-source GPT-4 moment" would unify the field. The ecosystem is fragmenting toward several credible open VLAs, each with different license regimes and different data coverage. Our current recommendation, which we develop in full in The Best Open-Source Robotics Stack 2026, is to design for portability between at least two VLAs rather than marry one.

6. What to watch in Q2 2026

• Bimanual benchmarks. Expect public benchmarks for two-arm manipulation to mature, driven by the ALOHA and OpenArm-style platforms. See our bimanual frontier analysis.
• Humanoid recall events. As install bases pass ~1,000 units per vendor, expect first serious field-reliability stories to break. Spare-parts networks will be put to the test.
• Closed-data leakage. Expect at least one closed VLA to either open its weights or face a credible open-weights competitor trained on a leaked recipe.
• Tactile sensing at scale. GelSight-style sensors are becoming cheap enough that large teleoperation operations are piloting them. The learning-side tooling is lagging — that gap will narrow.
• VLA-on-robot latency. Serving VLA policies at useful control rates on robot-side compute remains the least-solved infrastructure problem. We expect meaningful progress this quarter.

7. Practitioner recommendations

If you are setting a 90-day plan for Q2 2026, these are the choices we would make today:

• Data. Treat your teleoperation pipeline as a product with SLAs. Data quality ceiling-caps every downstream decision — see our quality checklist.
• Training. Default to LeRobot with OpenVLA or Octo, LoRA-tuned, on a single H100 or rented A100. Escalate to multi-GPU only when you have demonstrably outgrown single-GPU throughput.
• Hardware. Match the robot to the research agenda, not the agenda to the robot. Our store and leasing options cover both buy and rent. Browse robot comparisons before you commit.
• Deployment. Benchmark against your own evaluation suite, not a public leaderboard. Public benchmarks are necessary but not sufficient — see our catalog at /datasets/ and training recipes in tutorials.
• Team. Hire one data person for every two ML people. The binding constraint in Q2 will be data throughput, not model architecture.

8. Closing note

Q1 2026 was unglamorous in the best way. No single headline release rearranged the field. Instead, a stack that has been taking shape for two years became usable enough that small labs can now reproduce results that used to require an industrial research lab. The closed frontier kept moving, but the open frontier kept moving faster where it matters most for practitioners: reproducibility, cost, and the speed with which a new graduate student can go from laptop to working policy on a real robot.

If you want the full SVRC view on where this is heading, our future of the robot data market analysis is the companion piece. And if you are evaluating hardware or data services for the next quarter, get in touch — we run a small advisory practice for teams sizing their first real fleet.